ARTIFICIAL INTELLIGENCE FOR DETECTION OF INFLAMMATORY SACROILIITIS IN MAGNETIC RESONANCE IMAGING IN PATIENTS WITH AXIAL SPONDYLOARTHRITIS

J. Lee, S. Y. Kang, S. Lee, H. Kim, E. M. Koh, H. S. ChaSamsung Medical Center, Sungkyunkwan University School of Medicine, Department of Medicine, Seoul, Korea, Rep. of (South Korea) Health Insurance Review and Assessment Service, Healthcare Review and Assessment Committee, Wonjoo, Korea, Rep. of (South Korea)  Background Magnetic resonance imaging (MRI) is frequently used to evaluate active inflammatory sacroiliitis for early axial spondyloarthritis (axSpA) diagnosis. However, evaluation of sacroiliitis in MRI requires expertise because noninflammatory degenerative changes can mimic axSpA, and the semiquantitative diagnosis remains subject to significant variation. Artificial intelligence could function as assistance for inflammatory sacroiliitis detection. Objectives This study aimed to develop artificial intelligence for detecting inflammatory sacroiliitis for axSpA in MRI. Methods This retrospective study included MRI examinations of patients with clinical suspicion of axSpA collected at Samsung Medical Center between January 2010 and December 2021. Only the patients who performed short tau inversion recovery (STIR) MRI of the sacroiliac joints were included. Active inflammatory lesions consisting of bone marrow edema (BME) were identified independently by a rheumatologist and a radiologist using the Assessment of SpondyloArthritis international Society (ASAS) definition of MRI sacroiliitis. We propose a two-stage deep learning framework that combines a sacroiliac joints (SIJs) localization network with a BME classification network. First, the Faster R-CNN network extracts regions of interest (ROI) to localize the SIJs using whole MR images. Here, Maximum Intensity Projection (MIP) using three consecutive images of SIJs ROI is applied to enhance the low intensity of BME and consider the contextual information between images. Second, the VGG-19 network determines the presence of BME on individual MR images of localized SIJs ROIs with a resolution of 128x256. During the training process, we augmented the positive dataset 6-fold using blurring, contrast, noise, rotation, and sharpening because of the smaller number of data than the negative dataset. The prediction models were evaluated using 3-round, 3-fold cross-validation. The performance of BME classification was measured using accuracy and area under the receiver operating characteristic curve (AUC-ROC) curve. Results A total of 296 participants with 4,746 MRI images were included in the study. Inflammatory sacroiliitis was identified in 864 MRI images from 119 participants. The mean average intersection over unions (IoU) of ROIs to localize the SIJs was 0.742 for the right side and 0.744 for the left side. The mean accuracy and AUC-ROC of inflammatory sacroiliitis were 0.898 and 0.830 for image level and 0.801 and 0.827 for patient level, respectively. The confusion matrices of inflammatory sacroiliitis prediction are shown in Figure 1. In the original model, without using MIP and dataset augmentation, the mean accuracy and AUC-ROC were 0.867 and 0.609 for image level and 0.717 and 0.620 for individual level. Compared to the original model, improved performances of inflammatory sacroiliitis could be observed. Conclusion Artificial intelligence can detect inflammatory sacroiliitis for axSpA according to the ASAS definition in MRI. Image/graph: Acknowledgements This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. 2021M3H9A2097957 and no. 2021R1F1A1062148). The funders had no role in the study design; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. Disclosure of Interests None Declared. Keywords: Artificial intelligence, Imaging, Spondyloarthritis DOI: 10.1136/annrheumdis-2023-eular.3431Citation: , volume 82, supplement 1, year 2023, page 1728Session: Spondyloarthritis - clinical aspects (other than treatment) (Publication only)